[R] random interactions in lme

[Douglas Bates]

Jacob Michaelson wrote:
> 
> On Apr 24, 2005, at 8:52 AM, Douglas Bates wrote:
> 
>> Jacob Michaelson wrote:
>>
>>> Hi All,
>>> I'm taking an Experimental Design course this semester, and have 
>>> spent many long hours trying to coax the professor's SAS examples 
>>> into something that will work in R (I'd prefer that the things I 
>>> learn not be tied to a license).  It's been a long semester in that 
>>> regard.
>>> One thing that has really frustrated me is that lme has an extremely 
>>> counterintuitive way for specifying random terms.  I can usually 
>>> figure out how to express a single random term, but if there are 
>>> multiple terms or random interactions, the documentation available 
>>> just doesn't hold up.
>>> Here's an example: a split block (strip plot) design evaluated in SAS 
>>> with PROC MIXED (an excerpt of the model and random statements):
>>> model DryMatter = Compacting|Variety / outp = residuals ddfm = 
>>> satterthwaite;
>>> random Rep Rep*Compacting Rep*Variety;
>>> Now the fixed part of that model is easy enough in lme: 
>>> "DryMatter~Compacting*Variety"
>>> But I can't find anything that adequately explains how to simply add 
>>> the random terms to the model, ie "rep + rep:compacting + 
>>> rep:variety"; anything to do with random terms in lme seems to go off 
>>> about grouping factors, which just isn't intuitive for me.
>>
>>
>> The grouping factor is rep because the random effects are associated 
>> with the levels of rep.
>>
>> I don't always understand the SAS notation so you may need to help me 
>> out here.  Do you expect to get a single variance component estimate 
>> for Rep*Compacting and a single variance component for Rep*Variety?  
>> If so, you would specify the model in lmer by first creating factors 
>> for the interaction of Rep and Compacting and the interaction of Rep 
>> and Variety.
>>
>> dat$RepC <- with(dat, Rep:Compacting)[drop=TRUE]
>> dat$RepV <- with(dat, Rep:Variety)[drop=TRUE]
>> fm <- lmer(DryMatter ~ Compacting*Variety+(1|Rep)+(1|RepC)+(1|RepV), dat)
>>
>>
>>
> 
> Thanks for the prompt reply.  I tried what you suggested, here's what I 
> got:
> 
>  > turf.lme=lmer(dry_matter~compacting*variety+(1|rep)+(1|rc)+(1|rv), 
> turf.data)
> Error in lmer(dry_matter ~ compacting * variety + (1 | rep) + (1 | rc) +  :
>     entry 3 in matrix[9,2] has row 3 and column 2
> 
> Just to see what the problem was, I deleted the third random term, and 
> it didn't complain:
> 
>  > turf.lme=lmer(dry_matter~compacting*variety+(1|rep)+(1|rv), turf.data)
>  > anova(turf.lme)
> Analysis of Variance Table
>                    Df Sum Sq Mean Sq  Denom F value    Pr(>F)
> compacting          5 10.925   2.185 36.000  18.166  5.68e-09 ***
> variety             2  2.518   1.259 36.000  10.468 0.0002610 ***
> compacting:variety 10  6.042   0.604 36.000   5.023 0.0001461 ***
> 
> Now obviously this isn't a valid result since I need that third term; 
> but interestingly, it didn't matter which term I deleted, so long as 
> there were only two random terms.  Any ideas as to what the error 
> message is referring to?
> 
> Thanks for the help,
> 
> Jake Michaelson

Unfortunately, yes I do know what the error message is referring to - a 
condition that should not happen.  This is what Bill Venables would call 
an "infelicity" in the code and others with less tact than Bill might 
call a bug.